Performance comparison using different multilayer perceptron input–output formats to predict unsteady indoor temperature distribution

Abstract Computational fluid dynamics (CFD) is widely used to predict the indoor thermal environment; however, large time cost represents a significant disadvantage. Several deep learning approaches have been introduced to reduce prediction time in steady‐state predictions, though their feasibility under unsteady ones has yet to be investigated. Considering the flexibility of the multilayer perceptron (MLP) input–output format, this study compared the performance of two MLP input–output formats, MLP‐A (simultaneously outputting the values for all cells in a space in a single calculation run) and MLP‐B (outputting the values for a cell in each calculation run), when used to predict unsteady indoor temperature distribution in three scenarios: time interpolation, time extrapolation, and varying boundary conditions. The two considered input–output formats resulted in different prediction patterns in the time interpolation scenario: MLP‐B accurately predicted the spatiotemporal development of airflow compared to the CFD results, whereas MLP‐A did not. Both MLP‐A and MLP‐B performed poorly in the time extrapolation scenario but exhibited different error patterns. Finally, MLP‐A also generally provided a correct prediction of airflow development as well. This study contributes to an understanding of the prediction patterns provided by different MLP input–output formats for unsteady indoor airflow prediction.